AlphaFold Developed New Method Providing Physical Validity in 3 Style Models

AlphaFold 3 and similar artificial intelligence models predict the three-dimensional structure of protein complexes with surprising accuracy. But a significant part of the structures produced by these models is physically impossible: the chains overlap with each other, the ligament lengths and angles of leagues are broken, ring structures become irregular and stereomer orientations are reversed.

• The problem of current solutions: sampling routing with physical-based potential increases the cost of calculation and is not working on large devices due to memory load.
• New approach: The researchers developed a method that implements physical restrictions using projection technique during the removal. -Avantages: This technique makes the output of the model to be adapted to the physical validity without adding additional sampling cost.
• Application area: It can be applied directly for AlphaFold 3 style cofolding models in biomolecular structure prediction.

This study addresses the theoretical accuracy of AI models with real world restrictions. Providing physical validity, it has critical importance especially in medicine design and protein engineering. The new method increases the availability of outputs while maintaining the speed of models.

This development can ensure that the biomolecular simulation and the gap between AI is narrowened, the artificial intelligence-based exploration tools are more widely accepted in the industry.